Denoising and demosaicking of noisy bayer image using specialized networks

Abstract

Demosaicking is the process of recreating a full colored image from incomplete colored image. In digital camera, single sensor is used to make camera cost effective. The single sensor does not capture all colors for single pixel. To overcome this, color filter array (CFA) is used to obtain colored image from single sensor. The produced image from CFA is called mosaic image. There are many patterns of mosaic images but in this paper bayer pattern is considered. The mosaic image is corrupted by noise introduced by sensor or other hardware during capturing. Demosaicking on noisy mosaic image creates artifacts such as moiré and zippering. Some solutions have been proposed for denoising mosaic image but they are handcrafted solutions. In this thesis a solution is proposed to first denoise and then demosaick the image using machine learning. The mosaic image is denoised using CNN which is then demosaicked using residual learning strategy of single specialized network. One of the network is DHTN (deep high textured network) which is trained on textured images and second one is DSTN (deep smooth textured network) which is trained on smooth images. Experimental results shows that proposed solution gives better quality images than state-of-the-art schemes